Man I appreciate so much all the help you provided so far. I guess I'll keep digging. Would this meta scan cause Get or Scan on user tables to be very very slow too? That was actually was the problem I first noticed: from time to time a client Get or Scan would be extremely slow on my tables.
On Tue, Mar 1, 2011 at 10:38 AM, Jean-Daniel Cryans <[email protected]> wrote: > The store file looks fairly new, maybe it was just major compacted? > You could easily figure that out from the region server's log. > > Your problem is definitely a .META. region slowness. If scanning 8k > rows takes 30 minutes, scanning and updating 260 rows at the same time > probably isn't much faster. > > J-D > > On Tue, Mar 1, 2011 at 10:34 AM, Nanheng Wu <[email protected]> wrote: >> J-D, >> >> It looks like I don't have that many files: >> >> drwxr-xr-x - root supergroup 0 2011-02-28 23:21 >> /hbase/.META./1028785192 >> -rw-r--r-- 3 root supergroup 932 2011-02-03 19:23 >> /hbase/.META./1028785192/.regioninfo >> drwxr-xr-x - root supergroup 0 2011-03-01 17:13 >> /hbase/.META./1028785192/info >> -rw-r--r-- 3 root supergroup 11831330 2011-03-01 17:13 >> /hbase/.META./1028785192/info/6687606720393313750 >> drwxr-xr-x - root supergroup 0 2011-03-01 17:13 >> /hbase/.META./compaction.dir >> >> So you think compaction on .META. table on the RS could be the >> problem? while the log metaScanner is running I saw virtually no logs >> on that server though, I did see compactions once the meta scan ended >> and region close started to happen. I wonder what the cluster is doing >> while the meta scan is going. >> >> I would definitely upgrade in the near future, it's just that I am >> afraid upgrading won't guarantee a fix if I don't understand the >> nature of the problem more clearly. We also have another test cluster, >> running the same versions of everything on cheaper hardware and it >> doesn't have this problem. >> >> On Tue, Mar 1, 2011 at 9:52 AM, Jean-Daniel Cryans <[email protected]> >> wrote: >>> That long .META. scan might be the culprit, good catch! >>> >>> If you run this from the hadoop folder: >>> >>> bin/hadoop dfs -lsr /hbase/.META. >>> >>> Do you see a ton of files? Let's define "ton" as more than 16 files. >>> If so, and I see you have a very high number of regions, then you >>> could be hitting this problem: >>> https://issues.apache.org/jira/browse/HBASE-3499 >>> >>> In short, every few .META. change will result in a flush and it could >>> be that the region server hosting it isn't able to keep up with the >>> compactions. It's done so that since append isn't supported in your >>> hadoop that you won't lose too much data if the node dies, which is >>> really bad when it happens to .META. >>> >>> In your case it might not be so bad to set your .META.'s MEMSTORE_SIZE >>> back to the default value if you're not writing hence not splitting. >>> Running the script attached in that jira would do it (don't forget to >>> restart HBase after running it). >>> >>> Lastly, upgrading to HBase 0.90.1 and a hadoop that supports append >>> should be a priority. >>> >>> J-D >>> >>> On Tue, Mar 1, 2011 at 9:30 AM, Nanheng Wu <[email protected]> wrote: >>>> Hi J-D: >>>> >>>> I did the scan like you suggested but no splits came up. This kind >>>> of makes sense to me, since we write Hfiles using a MR job and then >>>> pretty much use HBase for read-only so maybe that's why the regions >>>> never got split? I think my slow disable problem is related to this >>>> RegionManager.metaScanner. I greped for that in the Masters log and I >>>> saw: >>>> 11/03/01 15:30:37 INFO master.BaseScanner: RegionManager.metaScanner >>>> scanning meta region {server: 10.146.14.38:60020, regionname: >>>> .META.,,1, startKey: <>} >>>> 11/03/01 15:55:40 INFO master.BaseScanner: RegionManager.metaScanner >>>> scan of 8296 row(s) of meta region {server: 10.146.14.38:60020, >>>> regionname: .META.,,1, startKey: <>} complete >>>> 11/03/01 15:55:40 INFO master.BaseScanner: RegionManager.metaScanner >>>> scanning meta region {server: 10.146.14.38:60020, regionname: >>>> .META.,,1, startKey: <>} >>>> Thread 27 (RegionManager.metaScanner): >>>> 11/03/01 16:21:04 INFO master.BaseScanner: RegionManager.metaScanner >>>> scan of 8296 row(s) of meta region {server: 10.146.14.38:60020, >>>> regionname: .META.,,1, startKey: <>} complete >>>> 11/03/01 16:21:04 INFO master.BaseScanner: RegionManager.metaScanner >>>> scanning meta region {server: 10.146.14.38:60020, regionname: >>>> .META.,,1, startKey: <>} >>>> Thread 27 (RegionManager.metaScanner): >>>> 11/03/01 16:46:07 INFO master.BaseScanner: RegionManager.metaScanner >>>> scan of 8298 row(s) of meta region {server: 10.146.14.38:60020, >>>> regionname: .META.,,1, startKey: <>} complete >>>> >>>> It looks like meta scanner is always running, each time taking 20-30 >>>> minutes. Here's what I did: >>>> >>>> 1) disable table -> at 16:28:31, according the log above there's a >>>> meta scan in progress. >>>> >>>> 2 ) the command hangs, no output in .META RS's log >>>> at 11/03/01 16:46:57 the client got 11/03/01 16:46:58 DEBUG >>>> zookeeper.ZooKeeperWrapper: Read ZNode /hbase/root-region-server got >>>> 10.146.14.38:60020 >>>> 11/03/01 16:46:58 DEBUG client.HConnectionManager$TableServers: Found >>>> ROOT at 10.146.14.38:60020 >>>> 11/03/01 16:46:58 DEBUG client.HConnectionManager$TableServers: Cached >>>> location for .META.,,1 is 10.146.14.38:60020 >>>> This corresponds exact to when the meta scan completed. .META RS >>>> started to spew some logs about closing regions. >>>> >>>> 3) Another meta scan started and the client hangs again until it >>>> finishes, and disable command returns.The whole thing took 2688.5930 >>>> seconds >>>> >>>> What kind of problem does it look like to you J-D? What does this >>>> background task metaScanner do? >>>> >>>> Thanks! >>>> >>>> >>>> >>>> On Mon, Feb 28, 2011 at 5:17 PM, Jean-Daniel Cryans <[email protected]> >>>> wrote: >>>>> I didn't get most of the information I was looking for... but I do >>>>> have something to work with. So I asked whether the metaScanner was >>>>> making progress or not, and from that log I think I see it does: >>>>> >>>>>> 11/03/01 00:47:30 DEBUG client.HConnectionManager$TableServers: >>>>>> Rowscanned=261, rowsOffline=3 >>>>>> 11/03/01 00:47:30 DEBUG client.HBaseAdmin: Sleep. Waiting for all >>>>>> regions to be disabled from myTable >>>>> >>>>> Only 3 rows are offline, whereas we expect all the region to be >>>>> offline. I expect you see those lines a lot? If so, one thing I'd like >>>>> to see is the result of this command: >>>>> >>>>> scan '.META.', {STARTROW => "myTable,,", LIMIT => 261} >>>>> >>>>> It's going to be big. Then grep in the result for the string SPLIT, >>>>> and please post back here the lines that match. >>>>> >>>>> J-D >>>>> >>>>> On Mon, Feb 28, 2011 at 5:04 PM, Nanheng Wu <[email protected]> wrote: >>>>>> I issued two more disable commands and looked at the .META RS. The >>>>>> last log entry looks like: >>>>>> >>>>>> 11/03/01 00:29:14 INFO regionserver.HLog: Roll >>>>>> /hbase/.logs/<RS_HOSTNAME>,60020,1298935751670/hlog >>>>>> .dat.1298935752660, entries=13527, calcsize=3839046, filesize=2560143. >>>>>> New hlog >>>>>> /hbase/.logs/<RS_HOSTNAME>,60020,1298935751670/hlog.dat.1298939354072 >>>>>> >>>>>> I issued the command at: 00:34:59, the client debug message has >>>>>> nothing but "11/03/01 00:34:59 DEBUG zookeeper.ClientCnxn: Got ping >>>>>> response for sessionid:0x12e6e938e9c000b after 90ms" until 00:47:19. >>>>>> Then all of a sudden it outputs the following: >>>>>> >>>>>> 0:47:28 DEBUG client.HConnectionManager$TableServers: Cache hit for >>>>>> row <> in tableName .META.: location server 10.146.14.38:60020, >>>>>> location region name .META.,,1 >>>>>> 11/03/01 00:47:28 DEBUG zookeeper.ZooKeeperWrapper: Read ZNode >>>>>> /hbase/root-region-server got 10.146.14.38:60020 >>>>>> 11/03/01 00:47:28 DEBUG client.HConnectionManager$TableServers: Found >>>>>> ROOT at 10.146.14.38:60020 >>>>>> 11/03/01 00:47:28 DEBUG client.HConnectionManager$TableServers: Cached >>>>>> location for .META.,,1 is 10.146.14.38:60020 >>>>>> 11/03/01 00:47:30 DEBUG client.HConnectionManager$TableServers: >>>>>> Rowscanned=261, rowsOffline=3 >>>>>> 11/03/01 00:47:30 DEBUG client.HBaseAdmin: Sleep. Waiting for all >>>>>> regions to be disabled from myTable >>>>>> 11/03/01 00:47:31 DEBUG client.HBaseAdmin: Wake. Waiting for all >>>>>> regions to be disabled from myTable >>>>>> >>>>>> Here's the full thread dump of the .META RS from 00:34:59 to 00:47:31 >>>>>> http://pastebin.com/Ywwx8afQ >>>>>> >>>>>> I hope I compiled some useful information here. I am pretty lost at >>>>>> this point, I would really appreciate your help! >>>>>> >>>>>> Thanks! >>>>>> >>>>>> >>>>>> >>>>>> On Mon, Feb 28, 2011 at 2:58 PM, Jean-Daniel Cryans >>>>>> <[email protected]> wrote: >>>>>>> Ok so no debug from the disable command? BTW you should output your GC >>>>>>> somewhere else like the .out file or a completely separate file. >>>>>>> >>>>>>> J-D >>>>>>> >>>>>>> On Mon, Feb 28, 2011 at 2:54 PM, Nanheng Wu <[email protected]> wrote: >>>>>>>> I see, so I should jstack the .META region. I'll do that. >>>>>>>> >>>>>>>> The master log pretty much looks like this: should I grep for >>>>>>>> something specific? >>>>>>>> >>>>>>>> 11/02/28 22:52:56 INFO master.BaseScanner: RegionManager.rootScanner >>>>>>>> scan of 1 row(s) of meta region complete >>>>>>>> 2011-02-28T22:52:57.138+0000: 263241.014: [GC 263241.014: [ParNew: >>>>>>>> 19136K->2112K(19136K), 0.0033350 secs] 378606K->36159 >>>>>>>> 6K(641944K) icms_dc=0 , 0.0034330 secs] [Times: user=0.02 sys=0.00, >>>>>>>> real=0.00 secs] >>>>>>>> 2011-02-28T22:53:04.956+0000: 263248.832: [GC 263248.832: [ParNew: >>>>>>>> 19068K->2112K(19136K), 0.0029690 secs] 378552K->36163 >>>>>>>> 0K(641944K) icms_dc=0 , 0.0030460 secs] [Times: user=0.02 sys=0.00, >>>>>>>> real=0.00 secs] >>>>>>>> 2011-02-28T22:53:12.664+0000: 263256.540: [GC 263256.541: [ParNew: >>>>>>>> 19136K->2112K(19136K), 0.0037690 secs] 378654K->36164 >>>>>>>> 7K(641944K) icms_dc=0 , 0.0038660 secs] [Times: user=0.01 sys=0.00, >>>>>>>> real=0.01 secs] >>>>>>>> 2011-02-28T22:53:20.608+0000: 263264.484: [GC 263264.484: [ParNew: >>>>>>>> 19136K->1810K(19136K), 0.0032600 secs] 378671K->36134 >>>>>>>> 6K(641944K) icms_dc=0 , 0.0033470 secs] [Times: user=0.02 sys=0.00, >>>>>>>> real=0.01 secs] >>>>>>>> 2011-02-28T22:53:28.278+0000: 263272.154: [GC 263272.154: [ParNew: >>>>>>>> 18834K->2112K(19136K), 0.0029870 secs] 378370K->36166 >>>>>>>> 1K(641944K) icms_dc=0 , 0.0030670 secs] [Times: user=0.01 sys=0.00, >>>>>>>> real=0.01 secs] >>>>>>>> 2011-02-28T22:53:35.919+0000: 263279.795: [GC 263279.795: [ParNew: >>>>>>>> 19136K->1747K(19136K), 0.0037090 secs] 378685K->36129 >>>>>>>> 8K(641944K) icms_dc=0 , 0.0037920 secs] [Times: user=0.02 sys=0.00, >>>>>>>> real=0.01 secs] >>>>>>>> 11/02/28 22:53:36 INFO master.ServerManager: 9 region servers, 0 dead, >>>>>>>> average load 760.2222222222222 >>>>>>>> >>>>>>>> >>>>>>>> On Mon, Feb 28, 2011 at 2:49 PM, Jean-Daniel Cryans >>>>>>>> <[email protected]> wrote: >>>>>>>>> Are you able to tell if that call in metaScanner is hanging or it's >>>>>>>>> making multiple calls to the .META. region? >>>>>>>>> >>>>>>>>> If former, then jstack the region server that hosts .META. and see >>>>>>>>> where it's blocked. >>>>>>>>> >>>>>>>>> if latter, then it means your .META. region is slow? Again, what's >>>>>>>>> going on on the RS that hosts .META.? >>>>>>>>> >>>>>>>>> Finally, what's the master's log like during that time? >>>>>>>>> >>>>>>>>> J-D >>>>>>>>> >>>>>>>>> On Mon, Feb 28, 2011 at 2:41 PM, Nanheng Wu <[email protected]> >>>>>>>>> wrote: >>>>>>>>>> J-D, >>>>>>>>>> >>>>>>>>>> Thanks so much for your help so far! I sent disable commands on 4 >>>>>>>>>> rather small tables and they got stuck for a long time again, so I >>>>>>>>>> took jstack of the master. From what I can tell, all disableTable >>>>>>>>>> calls are blocked by a meta scanner thread (sample log below). At the >>>>>>>>>> moment there were no other requests to the server at all. How should >>>>>>>>>> I >>>>>>>>>> investigate this further? If it helps, here's some background: I have >>>>>>>>>> several datasets, each of them is in a separate table; Our data >>>>>>>>>> pipeline produces a new version of each dataset everyday and only the >>>>>>>>>> lastest version should be used. This is how the data is loaded: for >>>>>>>>>> each dataset 1) run a MR job and outputs HFiles 2) call loadTable.rb >>>>>>>>>> to create a new table 3) disable and drop the previous version. As a >>>>>>>>>> result some calls to load table and drop table would overlap. Please >>>>>>>>>> let me know if something stands out to you as a potential culprit. >>>>>>>>>> Thanks! >>>>>>>>>> >>>>>>>>>> BTW, I am running Hadoop 0.20.2 with HBase 0.20.6 >>>>>>>>>> >>>>>>>>>> Thread 47 (IPC Server handler 13 on 60000): >>>>>>>>>> State: BLOCKED >>>>>>>>>> Blocked count: 3801 >>>>>>>>>> Waited count: 72719 >>>>>>>>>> Blocked on java.lang.Object@75ac522c >>>>>>>>>> Blocked by 27 (RegionManager.metaScanner) >>>>>>>>>> Stack: >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.master.TableOperation.process(TableOperation.java:154) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.master.HMaster.disableTable(HMaster.java:842) >>>>>>>>>> sun.reflect.GeneratedMethodAccessor9.invoke(Unknown Source) >>>>>>>>>> >>>>>>>>>> sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:25) >>>>>>>>>> java.lang.reflect.Method.invoke(Method.java:597) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.ipc.HBaseRPC$Server.call(HBaseRPC.java:657) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.ipc.HBaseServer$Handler.run(HBaseServer.java:915) >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> Thread 27 (RegionManager.metaScanner): >>>>>>>>>> State: WAITING >>>>>>>>>> Blocked count: 1526058 >>>>>>>>>> Waited count: 1488998 >>>>>>>>>> Waiting on org.apache.hadoop.hbase.ipc.HBaseClient$Call@4dd44ab0 >>>>>>>>>> Stack: >>>>>>>>>> java.lang.Object.wait(Native Method) >>>>>>>>>> java.lang.Object.wait(Object.java:485) >>>>>>>>>> org.apache.hadoop.hbase.ipc.HBaseClient.call(HBaseClient.java:722) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.ipc.HBaseRPC$Invoker.invoke(HBaseRPC.java:333) >>>>>>>>>> $Proxy1.get(Unknown Source) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.master.BaseScanner.checkAssigned(BaseScanner.java:543) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.master.BaseScanner.scanRegion(BaseScanner.java:192) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.master.MetaScanner.scanOneMetaRegion(MetaScanner.java:73) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.master.MetaScanner.maintenanceScan(MetaScanner.java:129) >>>>>>>>>> >>>>>>>>>> org.apache.hadoop.hbase.master.BaseScanner.chore(BaseScanner.java:153) >>>>>>>>>> org.apache.hadoop.hbase.Chore.run(Chore.java:68) >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> On Fri, Feb 25, 2011 at 10:23 AM, Jean-Daniel Cryans >>>>>>>>>> <[email protected]> wrote: >>>>>>>>>>> An hour to disable? That doesn't sound right at all :) >>>>>>>>>>> >>>>>>>>>>> I would approach this problem like I generally do with HBase issue, >>>>>>>>>>> first check the master log for any weirdness regarding my problem >>>>>>>>>>> (in >>>>>>>>>>> this case, grep for the table name). >>>>>>>>>>> >>>>>>>>>>> Then I would look the region server log(s) of the nodes that were >>>>>>>>>>> hosting regions from that table. You should see the steps taken to >>>>>>>>>>> disable the regions (starting to close, flush, region completely >>>>>>>>>>> closed). >>>>>>>>>>> >>>>>>>>>>> If you are able to do it while it's taking a very long time to >>>>>>>>>>> disable, try to jstack the process the seems to be hanging. >>>>>>>>>>> >>>>>>>>>>> Finally, like I said in another thread, there's a bug in 0.20.6 that >>>>>>>>>>> almost prevent disabling a table (or re-enabling) if any region >>>>>>>>>>> recently split and the parent wasn't cleaned yet from .META. and >>>>>>>>>>> that >>>>>>>>>>> is fixed in 0.90.1 >>>>>>>>>>> >>>>>>>>>>> J-D >>>>>>>>>>> >>>>>>>>>>> On Thu, Feb 24, 2011 at 11:37 PM, Nanheng Wu <[email protected]> >>>>>>>>>>> wrote: >>>>>>>>>>>> I think you are right, maybe in the long run I need to >>>>>>>>>>>> re-architect my >>>>>>>>>>>> system so that it doesn't need to create new and delete old tables >>>>>>>>>>>> all >>>>>>>>>>>> the time. In the short term I am having a really hard time with the >>>>>>>>>>>> disabling function, I ran a disable command on a very small table >>>>>>>>>>>> (probably dozen of MBs in size) and are no client using the >>>>>>>>>>>> cluster at >>>>>>>>>>>> all, and that took about 1 hour to complete! The weird thing is on >>>>>>>>>>>> the >>>>>>>>>>>> web UI only the region server carrying the META table has non-zero >>>>>>>>>>>> requests, all other RS have 0 requests the entire time. I would >>>>>>>>>>>> think >>>>>>>>>>>> they should get some request to flush the memstore at least. I *am* >>>>>>>>>>>> using the same RS nodes for some map reduce job at the time and top >>>>>>>>>>>> shows the memory usage is almost full on the META region. Would you >>>>>>>>>>>> have some idea of what I should investigate? >>>>>>>>>>>> Thanks so much. >>>>>>>>>>> >>>>>>>>>> >>>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >> >
